Learn and Adapt,or Perish: The Case of the F35 Lightning II

The Multi-Trillion Fighter Plane

Early 1990s, the US Marine Corps planned for a new fighter jet to replace the aging Harrier. The US Congress, worried about the prospect of massive costs associated with the development, procurement, operation, and support of different aircrafts for the various branches of the US military, sought to merge the demands of the Marine Corps with those of the US Airforce and US Navy (Congressional Research Service [CSR], 2020). The stated goal was to develop three variants of the same basic aircraft with high commonality to save on building and operations costs (Institute for Defense Analyses [IDA], 2010). Consequently, the Joint Advance Strike Technology (JAST) program was established, from which the F35 Lightning II fighter jet emerged.

The program run out of control, subject as it was to massive cost overruns, severe technical delays, and constant bickering between all actors involved. The lifetime costs of a single airframe may be up to $1.5 trillion (2015 prices, see Bender et al., 2015). As such, it has become the single most expensive project in US-history (Hughes, 2015). It fits in a long history of so-called megaprojects: governments and private actors engaging in projects that promise the technological sublime (Frick, 2008) but that in most cases tend to run out of control (Flyvbjerg, 2014; Flyvbjerg et al., 2003). With so much evidence about the troublesome nature of megaprojects, one can be forgiven for asking “haven’t they learnt?” This is especially the case with the F35, which was preceded by the equally troublesome (but less expensive) F22A Raptor program of the US Air Force. While the obvious answer to that question seems a resounding “no,” a coevolutionary perspective may generate a different kind of answer that aligns more closely to the nature of such complex processes.

Our contribution to this special issue will assess to what extent a (co)evolutionary view on learning may enhance our understanding of learning in governance processes. Generally speaking, evolution and co-evolution explain the emergence of speciation through variation, selection, and retention under selection pressure to establish fit with the environment. Such a fit would ensure long-term survival. We therefore ask the question: in what ways do actors learn to improve their chances of long-term survival in complex governance processes? To answer this question, we first introduce the concept of coevolution and its relationship to learning in Section 2. From this, we derive a fitness landscape model of collective decision making that structures the roles, relationships, preferences, and behaviors of actors, and the extent to which they are successful in achieving long-term survival, in governance processes (Section 3). In Section 4, this model is applied to the JAST/JSF-program, followed by the analysis in what ways the involved actors have learnt to increase their chances of survival in Section 5. We conclude that actors that have learnt the best how to survive the changing environment did so at the cost of a dysfunctional technological program that has cost billions of tax money across the globe. We reflect on different modes of learning in relationship to coevolution.

Learning in Coevolution

Broadly speaking, evolution is the change of characteristics in systems—species, actors, technologies, institutions, etc.—under selection pressure and in duration. Arguably, evolution is reciprocal as a change of properties under environmental selection pressure may cause selection pressure on the environment in return. Adaptation concerns the selection of those characteristics within a given system to suit selection pressures; coevolution expresses; and maps that this adaptation causes the environment to change, too (Gerrits, 2008; Gerrits et al., 2009). Owing to the nature of selection in a coevolutionary interaction, certain systems may benefit from the interaction while the other (coevolving) system is not, negatively or positively affected. In other words, coevolution can be beneficial or detrimental, or all shades of gray in between, for either or both populations (Gerrits & Marks, 2017; Odum, 2004).

Coevolution, as originally coined by evolutionary biologists Ehrlich and Raven (1964), has spilled-over into the social sciences as a way of conceptualizing reciprocity between different types of systems, and between agents within systems (e.g., Norgaard, 1994, 1995; see also elsewhere in this thematic issue). The concept has proven to feature explanatory value (Gual & Norgaard, 2010; Hird, 2010; Holling, 2001; Weisz & Clark, 2011), in particular when it comes to socio-ecological interaction where it has been studied extensively. It spawned an entire school of research, with applications covering diverse topics such as the exploitation of (rain)forests (e.g., Lambin & Meyfroidt, 2010; Norgaard, 1994), water resource management (e.g., Kallis, 2010; Van Staveren et al., 2018), flood risk management (e.g., Tempels, 2017), development of estuaries (e.g., Gerrits, 2008, 2011), agriculture and its land use (e.g., Moreno-Peñaranda & Kallis, 2010), and hydrology (e.g., M. R. Sanderson, 2018). Coevolution also speaks to matters of Public Administration. For example Norgaard (1994, 1995) and Rammel et al. (2007) incorporate institutional analysis and individual behavior, while Gerrits (2008, 2011) and Hood (2017) focus on the dynamics of policy making in relationship to governing of ecosystems. Key texts include Teisman et al. (2009), Van Assche et al. (2014), and Beunen et al. (2015).

As with many other concepts that have travelled between scientific domains, coevolution has been interpreted and used in a variety of ways in target domains. As mentioned in the introduction to this special issue, some of these interpretations and uses are complementary where others are incompatible. A principal distinction must be drawn between the general, conceptual idea of things changing in tandem on the one hand, and the analysis of the mechanisms of such tandems on the other hand (S. Sanderson, 1990). While the first articulates the general logic on a conceptual level without much further analytical depth, the second attempts to map and explain the causal dependencies that drive the coevolution between, for example, systems or actors. A second distinction must be drawn between coevolution as a concept that explains reciprocity—indifferent of the outcomes—and coevolution as a normative value, that is, coevolution as something that is inherently good. While the first departs from a systems perspective where certain behaviors are selected for given the circumstances, the latter starts from the perspective of individual agents having purposeful action. Key point is the difference between intentional adaptation through learning—where actors search for options and learn what works and what does not work in a given situation (e.g., Siggelkow, 2002; Siggelkow & Rivkin, 2006)—and blind selection where certain environmental pressures select for characteristics regardless of learning effects (e.g., Gavrilets, 2003). Given these two distinctions we follow the analytical approach from a systems perspective where behaviors are selected reciprocally and without teleological assumptions. Hence, we discard the normative route.

Environmental pressures select for particular properties that then survive in the long run. This requires some additional explanation when applied to the social realm. Survival is achieved when something—for example, an actor—has adapted to selection pressures such that it fits with its environment and can continue to exist. The notion of fitness stretches beyond a mere “survival” when it comes to the topics central in Public Administration. Lasswell’s (1936) “who gets what, when, and how” is the key to fitness. Fitness may be achieved in direct and indirect ways. It is therefore better to speak of inclusive fitness, which denotes the effect of an actor’s actions on its own fitness and that of others directly related to this actor (Grafen, 2006; Hamilton, 1964), which may be reciprocal. This also takes into account that actors may do things that are directly beneficial to others but not themselves, at least not in the short run (Birch, 2016) In other words, inclusive fitness is the extent to which actors reach their goals given the dynamic environment they are in, where other actors are also trying to reach their goals. There are many ways in which actors can obtain fitness in governance processes; for example, one may accept a short-term loss in exchange for long-term survival. An importance difference with biology is that human actors can try to forecast, plan, anticipate, and reflect on the effects of actions undertaken of one’s own survival chances. This is where learning comes into focus.

Learning in coevolutionary processes may be analyzed at various levels—from the individual, day-to-day interactions, to system-wide perspectives. From a system’s perspective—the one adopted in this paper—actors receive feedback about what strategies and behaviors contribute to an actor’s fitness (and therefore survival) given the environmental conditions in which they have to survive. That is: learning through reflexivity (Newig et al., 2016). Comparison (cf. Dunlop, 2017) of strategies deployed and behaviors shown by various actors in the same arena in relationship to their ability to obtain relative fitness helps actors to learn about what is more effective in certain situations.

A change in environmental conditions may render the adopted behaviors obsolete, as such necessitating a new learning effort through variation of behaviors (searching for the best way forward) and subsequent selection and retention of the best fitting strategies and goals. Arguably, those who survive in the long run are the ones who learnt best how to deal with the selection pressures in coevolution, that is, those actors have the highest fitness as they achieved most of their goals. The necessity to learn is driven by a desire for survival. This is not to say that learning is always an intentional strategy directly focused on reaching a better fit. Indeed, there is much uncertainty about what a better fit may constitute in the first place. As such, actors will experiment with strategies and behaviors in an attempt to get a feeling for what the environment demands from them (cf. learning through experimentation). This includes the possibility of getting it entirely wrong and losing evolutionary advantages.

Fitness Landscapes as Mapping Device

To study the evolutionary dynamics outlined above in governance processes, we deploy a fitness landscape model, which is a common model for analysing such dynamics (Gavrilets, 1998, 2004, 2014), and that suits the study of collective decision-making equally well (Gavrilets & Richerson, 2017; Gerrits & Marks, 2017; Marks et al., 2019), in order to track the strategies, behaviors, goals, and fitness of actors in a collective decision-making process. We will explain the model below. As per Marks et al. (2019) and Weisberg (2007), this concerns both the semantic and syntactic structure of collective decision-making, that is: the conceptualization of the process and the definition of the model go hand-in-hand.

In this conceptualization, actors coordinated repeatedly over a specific issue in multiple governance arenas. Repeated coordination over time ensures learning effects as actors receive feedback about what works and what does not. This repeated coordination is structured into a series of connected events that together form a lineage. Following Abbott (2001), descriptions that bring together the contextual information lead the researchers to select and assign events to a lineage. As per the rounds model (Teisman, 2000), major events may significantly impact the issue, demarcate the start or end of such a process, or new actors may enter or depart. Each string of events between those major events is captured in a fitness field populated by the actors active in that field. The aforementioned “who gets what, when, and how” brings in the actors involved and their related goals. A lineage ends when the collective decision about the issue is reached. There can be multiple lineages—for example, technical, market-oriented, or political—that diverge, converge, or intersect. Owing to the coevolutionary nature of all (complex) governance decision-making processes, lineages represent different aspects of those decision-making issues that may influence reciprocally (Gerrits & Marks, 2017).For example, decision-making about financing fighter jets is intertwined with international and diplomatic coalition-forming.

Actors hold problem definitions of the issue as they perceive and frame it in their own actor-specific ways, and they formulate their own solution definitions as they have specific ideas how the issue should be solved (i.e., framing of an issue, see Fischer, 1998; Fischer & Forester, 1993; Rein & Schön, 1996). These combine in problem and solution definitions (PSDs) that express the ways in which actors think of the issue at hand. The space of possible problem and solution definitions is defined by the number of PSDs observed empirically in a given fitness field, where

every actor i has P S D i = a c t u a l # e l e m e n t s P S D i m a x i m u m # e l e m e n t s P S D , where P S D i ∈ [ 0 , 1 ] .

As all actors can hold multiple PSDs about the issue the chance of realizing certain PSDs is strongly related to the PSDs of the other actors. The interdependence of the actors is also prevalent in resource dependencies (e.g., finance, knowledge, etc.) and formal responsibilities (e.g., as defined in a constitution, rules and regulations, etc.). How well actors are connected in a fitness field is measured by the number of actual links that an actor in a network has as a rate of the number of possible links (c_score, cf. density in social network analysis, Tichy et al., 1979), thus

c s c o r e i ( t 0 ) = a c t u a l c o n n e c t i o n s i m a x i m u m # c o n n e c t i o n s , where c _ s c o r e i ∈ [ 0 , 1 ] .

Not all actors will reach their goals, as is common in collective decision-making. While failure to reach a goal does not end an actor’s existence instantly, actors do need to succeed in (partial) realization of their goals if they want to survive in the long run. Actors may deploy two principal strategies to realize the most of their PSDs. First, they can seek alignment with certain actors, and increase distances with others if they think this will contribute to goal attainment for a given issue based on their perceptions, strategies, perceived distances, and motives of those other actors (March, 1994; Marks & Gerrits, 2018). Second, they learn which strategies and alliances may work better through comparison and reflection (Dunlop, 2017). Actors will seek to align their PSDs with similar PSDs (e.g., package deals) while distancing from dissimilar PSDs (e.g., non-negotiable differences). This process of reciprocal informing means that a qualitative adjustment based on content (similar elements of the PSD) enters, attributing weight (w) to the c_score for every actor. In other words,

c s c o r e i ( t ) = w i × a c t u a l c o n n e c t i o n s i m a x i m u m # c o n n e c t i o n s ; where w i = # a c t u a l s i m i l a r e l e m e n t s P S D i m a x i m u m # s i m i l a r e l e m e n t s P S D , and w i ^ c _ s c o r e i ∈ [ 0 , 1 ] .

This allows mapping the actors involved in a certain issue relative to each other and over time. Throughout the process, actors may change in either or both dimensions, or stay put if they feel that this is the best strategy for the moment. Naturally, the field will change as soon as actors start moving (internal to the field) and/or major events take place (external to the field). As such, it may be that an actor doesn’t intend to change in the principal dimensions but still finds itself in a different place because others have changed theirs. The movements of all actors together impact the likelihood of goal attainment. Goal attainment, which we see as a proxy for fitness and therefore long-term survival is expressed as fitness (f) in the model, where,

f i = v ( P S D i , c _ s c o r e i ) for each actor I , where f i ∈ [ 0 , 1 ] .

where v is the value attributed in retrospect by the researchers using case-based knowledge. The model does not assume an a priori fixed relationship between v(PSD_i, c_score_i) and f_i. Exactly which combinations of PSD and c_score will lead to goal attainment or loss is something that needs to be investigated empirically. Given sufficient data, one will be able to tease out persistent configurations of all three dimensions over time.

Strategies in governance processes concern the combinations of problem and solution definitions, and the extent to which actors are connected to actors. As mentioned above, actors considered in the analysis of governance have agency, that is, the ability to anticipate, forecast, plan, respond, learn, and reflect. Actors will learn about what strategies work and which not via changes in their fitness value. Over time, strategies contributing to fitness are retained, and strategies not contributing to fitness are selected against. Arguably, the model is an abstraction that ignores details at the level of individual actors, but it allows for unambiguous assessment that uncovers persistent patterns between content, connection, and fitness.

Data and Methods

We use the fitness landscape model to analyze the collective decision-making process leading up to the acquisition of the F35 Lightning II fighter jet to demonstrate how actors learn the strategies needed to ensure long-term survival. To this end, we traced the decision-making process between 1996 and 2016 through 492 articles published on the topic in newspapers from the USA, Canada, Australia, and the United Kingdom. In addition, we used 10 policy documents from the US government; particularly the 2009 root cause report issued by the Office of the Under Secretary of Defense, the Selected Acquisition Report (SAR) issued by the Department of Defense (DoD), and the 2020 updated F-35 JSF program report by the Congressional Research Service. ¹

All articles were chronologically structured into a database and then coded to reconstruct how different events evolved over time. This made it possible to select the significant events to form the lineages. The dynamics of those events were fleshed out using the government documents. This enabled us to reconstruct the connections between actors, their PSDs, the decisions they took, and the context of those decisions; that is, the main actors and the main events per lineage and related fitness fields are identified. After the DoD announced it wanted three versions of a new hi-tech fighter jet (1994), two decades of subsequent decision-making can be divided into three main lineages:

Decisions and debates about who may build the jet at what cost: that is, all events of the competition between potential manufacturers.

Decisions about what the jet should be able to do given budget constraints: that is, all events related to the actual design and building of the aircraft.

Decisions about attracting other countries into the program: that is, events related to the building of an international alliance around the program.

Owing to the coevolutionary nature of decision-making processes certain actors, decision or outcomes in one field or lineage may intersect with another related field in another lineage. The three lineages and their coevolutionary links are depicted in Figure 1, followed by a short description in the following section.

OPEN IN VIEWER

Figure 1.

The three lineages comprising the JAST-JSF study, including the moments in time when these lineages were linked reciprocally (vertical lines in gray).

Discussion and Conclusions

While the future will tell if the F35 Lightning II was worth the trouble, the current situation shows all the hallmarks of a botched megaproject. Notwithstanding the intuitive but ultimately naïve response that it appears as if none of the actors learnt from the pitfalls of such megaprojects (e.g., Flyvbjerg et al., 2003), the evolutionary view shows that the outcome is the result of actors learning how to safeguard their own long-term survival—in terms of contracts, investments, employment, obtaining jet fighters, and so on. These actors learn to align with the DoD by connecting and incorporating elements of DoD’s PSD. LM learns to repeatedly align with DoD PSDs and also to incorporate PSDs of Members of Congress to ensure fitness. In doing this, LM created the certainty that even if they encountered problems in the production or technical specifications, funding would be guaranteed because Members of Congress would protect their workforce and getting the funds necessary (notice the coevolutionary relation between PSDs and connections between lineage 1 and 2 here). Both Member of Congress and LM are learning through updating their prior approaches in the light of the reciprocally updated information of each other (Newig et al., 2016). Similarly other nations connected their faith to the DoD; wanting to safeguard their homeland jobs and income generated by the production or maintenance of the fighter jet constantly make these nations align with DoD’s PSD. Again, an affordable fighter jet was only part of the respective PSDs. Once sunk costs were so high no viable alternative remained.

What do the results of this study show about learning in coevolutionary governance processes? As mentioned in the introduction of the special issue, coevolution is a given, that is, something that will take place regardless of intentions of actors. As actors change positions in the policy arena, other actors will have to find out if they want to move along or search for alternative routes through the shifting landscape. The drive for survival is expressed system-wide in terms of alignment and divergence, propagation of substantive and network diversity, and self-organized entrapment. These three dynamics are reinforced because of two principal learning processes: experimentation with, and evaluation of (combinations of) PSD and connections in relationship to perceived fitness gains and losses; and comparison of (combinations of) PSD and connections between actors to identify successful combinations that may be adopted in an attempt to improve one’s fitness. The study shows that the initial stage of such governance processes is pivotal in how things play out: a high diversity of problem and solution definitions and coalitions converges toward the main set of PSD and network structure over time. Actors can opt to align further with other actors, or opt for a go-alone strategy, depending on the outcome of the evaluation. Alignment implies further exposure to alternative problem and solution definitions and strategies, often fueled by input from external experts and practices observed elsewhere. A positive feedback loop driving further diversification may emerge. Likewise, too much alignment can lead to self-organized entrapment. Results from all movements feed back into the actors’ evaluation of their own fitness in contrast to the fitness of others. The process is summarized in Figure 2 below.

OPEN IN VIEWER

Figure 2.

Primary learning mechanisms as seen from the coevolutionary framework used in this paper. Any actor will compare its own fitness against the fitness of others elsewhere in the landscape. Observing that improvement of fitness is possible, the actor will compare its own PSD and c_score against that of others to determine what strategies (expressed in the c_score) and goals (expressed in PSD) are more likely to bring improvement of fitness (Van Assche et al., 2021, p. 13). Subsequently, actors may align with other actors in terms of c_score and PSD. Alignment implies an exposure to alternative ideas and strategies (Van Assche et al., 2021, p. 13), that is, higher diversity. This is also fuelled by exposure to experts and practices from elsewhere (a positive feedback loop between exposure and learning to diversify may emerge; Van Assche et al., 2021, pp. 13, 14). Excessive alignment leads to self-organized entrapment. Non-alignment, alignment on the basis of diversification, and self-organized entrapment all feed back into actor’s evaluation of its own fitness vis-à-vis the fitness of others.

Naturally, this process is not cast in concrete. In particular, we expect initial stages to be characterized by diversification and experimentation while later stages may be more dominated by convergence. Over time, the space-of-possibilities becomes delineated and more restricted by those serial learning experiences (Newig et al., 2016, p. 354; Van Assche et al., 2021, p. 13) about what works and what doesn’t. If a certain combination turns out to be successful, it is very likely that this combination is repeated elsewhere in time and place, as was shown in this study.

The failures of certain actors (who may even disappear from the process altogether) is as instructive as is the characteristics of the successful actor(s). And while the most fitting combinations will be clear in hindsight, one has to understand and appreciate that actors find themselves in considerably more uncertain conditions—especially in the early stages—that requires from them to be reflexive and adaptive. It confirms that history and the present are the most important drivers behind learning (Dunlop & Radaelli, 2013; Newig et al., 2016; Van Assche et al., 2021). Importantly, it is not only their own intentions that matter—even DoD didn’t control every field—but above all the movements of others in the field, as well as the impact of external events they have to deal with. This perspective puts much emphasis on adaptive behavior in the face of coevolving processes. The larger objective—a functional and affordable fighter jet—seems to be of secondary importance in the light of the struggle for survival. The F35 case is far from unique, in the sense that actors went for those combinations that were most likely to be successful. It is not uncommon for the military to ask for future-proof (speculative) technology, or for members of Congress to direct funds to their states. The experience that these goals often contributed to long-term survival (of the military and of members of Congress) means that they are taken into consideration in the next project. In other words, the space of possibilities is also delineated by earlier learning experiences. As such, it is no surprise that actors are willing to embark on yet another megaproject as they have learnt that this is an opportunity for survival.

The study presented here contributes to the questions raised in the special issue from the vantage point of governance as an ecosystem. This view allows us to consider the very long-time spans inherent to coevolutionary processes. The study could be criticized for taking the view that survival is the best indicator for learning, as this may ignore individual learning in coevolution, as such not addressing the dialectic learning mentioned elsewhere in this special issue—even though one can sense from the results of our study that little actual reflection on the nature of the governance process took place. We understand that criticism but would like to highlight that the perspective selected in this paper demonstrates that, ultimately, selection is blind. That implies that authors can (and will) try to achieve fitness gains but always against a background of uncertainty. The outcomes emerge—literally—from the many micro-interactions and distributed learning effects that have taken place. The structural mechanisms of variation, selection, retention, and fitness underneath coevolution explain why this is so. As such, they highlight why actors went ahead despite the mounting problems. The model and method presented here can unearth those mechanisms and may complement a micro-level analysis.